The best way to use ChatGPT for work is to give it a defined deliverable, reliable source material, clear boundaries and a test for whether the result is acceptable. It can then help research, draft, compare, calculate, organize and create. A person should still verify the evidence and remain responsible for every consequential decision or external action.
That is a different approach from opening an empty chat and asking ChatGPT to “help with work.” A vague request invites a generic answer. A repeatable workflow specifies the inputs, the job to perform, the output format and the checks that must happen before anyone uses the result.
This guide contains 25 such workflows. Each one explains:
- what the workflow produces;
- what information or files to provide;
- which ChatGPT feature fits the job;
- an exact prompt you can adapt;
- how to check the result;
- which sensitive information to withhold or redact.
If you first want to understand the product, models and features, read our complete guide to ChatGPT. If you need to decide which subscription supports your intended workload, compare the plans in our ChatGPT pricing guide. This page does not repeat either subject. It focuses on putting ChatGPT to work.
Features and availability were checked on July 24, 2026. ChatGPT changes frequently, and access can depend on the plan, device, region and workspace settings.
25 ChatGPT work examples at a glance
| # | Workflow | Best starting point | Main output |
| 1 | Draft an email reply from a thread | Chat or Work | Reviewed email draft |
| 2 | Turn rough notes into a decision brief | Chat | One-page brief |
| 3 | Prepare for a meeting | Project or Work | Meeting briefing pack |
| 4 | Convert a meeting transcript into actions | Record, Chat or Work | Decision and action log |
| 5 | Produce a weekly status report | Project or Work | Consistent progress report |
| 6 | Research a business question | Deep Research | Cited research report |
| 7 | Build a competitor evidence matrix | Deep Research | Comparable fact table |
| 8 | Synthesize interviews and surveys | Chat or Data Analysis | Themes and evidence |
| 9 | Write a decision memo | Chat or Work | Options and recommendation |
| 10 | Compare two document versions | Chat or Work | Change and impact table |
| 11 | Review a document against requirements | Work | Compliance matrix |
| 12 | Extract obligations and deadlines | Chat or Work | Obligation register |
| 13 | Analyze a spreadsheet | Data Analysis | Findings, tables and charts |
| 14 | Clean and restructure a dataset | Data Analysis or Work | Clean file and change log |
| 15 | Build a forecast and scenarios | Data Analysis | Assumption-based model |
| 16 | Create a presentation from sources | Work | Editable slide deck |
| 17 | Build a project plan | Chat or Work | Milestones and dependencies |
| 18 | Create a risk register | Chat or Work | Prioritized risks and controls |
| 19 | Turn a process into an SOP | Work | Step-by-step procedure |
| 20 | Draft a marketing campaign brief | Chat or Work | Campaign plan |
| 21 | Repurpose approved content | Chat or Work | Channel-specific assets |
| 22 | Draft policy-grounded support replies | Chat or Work | Customer reply draft |
| 23 | Find patterns in support tickets | Data Analysis | Root-cause analysis |
| 24 | Standardize a hiring process | Chat or Work | Scorecard and interview kit |
| 25 | Schedule a recurring monitor | Scheduled Tasks | Briefing or change alert |
ChatGPT Work versus using ChatGPT at work
“Using ChatGPT at work” describes any professional use of ChatGPT. Someone drafting an email in regular Chat is using ChatGPT at work.
ChatGPT Work, with a capital W, is a named product experience for longer and more involved assignments. It can work across supplied files and approved connected apps, carry out multiple steps and create finished documents, spreadsheets, presentations, reports and other deliverables. You can follow its progress, answer questions, change direction and approve important actions.
The practical distinction is:
- use Chat for a short, interactive task;
- use Search for a quick current fact with links;
- use Deep Research when the answer requires several sources and a documented report;
- use Data Analysis for calculations, tables, charts and structured files;
- use Work when the task has several steps or should end in a finished artifact;
- use Scheduled Tasks when the work must happen later, repeat or wait for a change.
ChatGPT Work is rolling out across supported paid plans and surfaces, but availability can still depend on plan and workspace controls. Free and Go users can still perform many of the workflows below through regular Chat, files, Search and other included features, although limits and tool access differ. Our ChatGPT Work launch coverage explains the introduction of the named experience; this guide covers the broader workplace method.
Before using any of these workflows
Do not treat access to ChatGPT as permission to upload every file you can open.
First check your employer’s AI, security, confidentiality, records and acceptable-use policies. Use an approved account or workspace. Share only the minimum information the task requires, remove irrelevant personal data, and never paste passwords, API keys, authentication codes or private signing credentials into a prompt.
For personal ChatGPT accounts, data controls determine whether new conversations may be used to improve models. OpenAI says business workspace data is excluded from training by default, but that does not remove your own legal, contractual or organizational duties. Retention, app permissions, administrator controls and the sensitivity of the underlying information still matter.
Every workflow below therefore includes a data boundary. Treat it as a minimum safeguard, not a substitute for your organization’s rules.
For stronger prompts in general, see our complete ChatGPT prompts guide. The prompts here are deliberately task-specific.
Communication, meetings and reporting
1. Draft an email reply from an existing thread
Use ChatGPT to produce a first draft when the facts already exist in an email thread but the response needs to be clearer, shorter or better structured.
Provide: The relevant messages, the recipient’s role, the outcome you want, any commitments you are authorized to make, the desired tone and the deadline.
Use: Regular Chat for one reply. Use Work or an approved email app when ChatGPT must find the thread or save a draft. Keep sending as a separate human-approved action.
Reusable prompt:
Draft a reply to the email thread below.
Goal: [what this reply must achieve]
Recipient: [role and relationship]
Tone: [direct, warm, formal, concise, etc.]
Required facts: [facts that must appear]
Authorized commitments: [what I may promise]
Forbidden commitments: [what I may not promise]
Desired next step: [action, owner and date]
Use only facts in the thread and the notes above. Do not invent a date,
attachment, policy or commitment. If a necessary fact is missing, list the
missing fact before drafting.
Output:
1. Suggested subject line
2. Email of no more than [number] words
3. A fact-check list of names, dates, figures and commitments I must verify
Do not send the email.
[Paste the relevant thread]
Verify: Check every name, date, amount, attachment reference, promise and requested action against the original thread. Read the email once from the recipient’s perspective before sending it yourself.
Data boundary: Remove unrelated messages, personal addresses, signatures and third-party details. Never include passwords, access links, payment-card data or an entire mailbox when one thread is enough.
2. Turn rough notes into a decision-ready brief
ChatGPT can impose structure on fragmented notes without pretending that missing evidence exists. The result is useful for an executive update, internal proposal or handover.
Provide: Raw notes, the intended reader, the decision or discussion the brief should support, a maximum length and any terminology that must remain unchanged.
Use: Regular Chat for a short brief. Use a Project when the same audience, format and background material will be reused.
Reusable prompt:
Convert the notes below into a one-page decision brief for [audience].
The reader needs to decide or understand: [decision/question]
Maximum length: [word count]
Required terminology: [terms]
Structure the brief as:
- Situation
- What the evidence says
- Decision or input needed
- Options, if the notes support them
- Risks and dependencies
- Open questions
- Recommended next step
Do not add outside facts. Distinguish confirmed facts, estimates and opinions.
If two notes conflict, show the conflict instead of resolving it silently.
End with a short list called "Claims to verify."
[Paste notes]
Verify: Trace every material claim back to a supplied note. Confirm that opinions have not been converted into facts and that the requested decision is explicit.
Data boundary: Remove private comments about colleagues, unneeded customer details and unapproved financial or strategic information. Replace names with roles when identity does not affect the task.
3. Prepare a meeting briefing pack
A briefing pack is more useful than a generic agenda. It should tell the attendee what the meeting is for, what has already happened and which questions or decisions require attention.
Provide: The agenda, previous minutes, relevant emails or project documents, attendee roles, the meeting objective and the authority you have in the room.
Use: A Project keeps recurring meeting material and instructions together. Use Work if ChatGPT must assemble several sources into a finished briefing document. An approved calendar or email app can retrieve context when available.
Reusable prompt:
Create a briefing pack for this meeting.
Meeting: [name, date and duration]
My role: [role]
Primary objective: [objective]
Decisions I can make: [authority]
Decisions that require escalation: [limits]
Using only the supplied agenda and source material, produce:
1. A 100-word executive briefing
2. The three outcomes I should pursue
3. A timeline of relevant prior decisions
4. Questions to ask, ordered by importance
5. Likely points of disagreement and the evidence relevant to each
6. Decisions required in the meeting
7. A note-taking table with Decision, Owner, Deadline and Evidence
Do not invent information about attendees. Mark missing context as "unknown."
[Attach or paste agenda, minutes and source material]
Verify: Confirm attendee roles, current project status and which decisions are actually open. Remove any speculative description of a person’s motives.
Data boundary: Do not include private calendar details, personal evaluations, medical information, confidential HR discussions or unrelated email content.
4. Convert a meeting transcript into decisions and actions
ChatGPT can turn a transcript into usable follow-up material, but transcription errors can change names, numbers and commitments. The transcript remains evidence to inspect, not an infallible record.
Provide: A transcript or approved recording output, the agenda, the attendee list and any required action-log format.
Use: ChatGPT Record can transcribe and summarize meetings where available. As of the review date, Record is available in eligible workspaces through the macOS desktop app. Otherwise, upload an authorized transcript to Chat or Work.
Reusable prompt:
Analyze this meeting transcript. Do not create an owner, deadline or decision
unless the transcript supports it.
Produce:
1. Decisions made, with the transcript passage or timestamp supporting each
2. Action items in a table: Action | Owner | Deadline | Evidence | Confidence
3. Questions that remained unresolved
4. Risks or dependencies raised
5. Statements that sound like proposals but were not accepted
6. A follow-up email of no more than 180 words
If a name, number or date is uncertain, quote the ambiguous wording and mark it
for review. Do not treat general discussion as agreement.
[Attach transcript and agenda]
Verify: Replay or reread every important decision, owner, deadline and numerical claim. Ask participants to correct the action log rather than presenting an AI summary as the official record without review.
Data boundary: Obtain the legally required consent before recording people. Do not upload privileged meetings, disciplinary discussions, health information or trade secrets unless the recording and workspace are explicitly approved for that use.
5. Produce a consistent weekly status report
The valuable part of an AI-generated status report is not polished prose. It is consistent comparison across the reporting period: what changed, what slipped, what is blocked and what needs a decision.
Provide: The previous report, current task log, milestone plan, agreed metrics, decisions and the exact reporting dates.
Use: A Project is useful for recurring reports because it keeps the format and background together. Work can assemble the source pack and create a finished document.
Reusable prompt:
Create the weekly status report for [project] covering [start date] through
[end date].
Use the previous report only as a format and baseline. Use the current source
pack for this week's facts.
Required sections:
- Overall status: Green, Amber or Red, with a one-sentence evidence-based reason
- Completed this period
- Planned for next period
- Milestones: planned date, current forecast and variance
- Metrics: current value, prior value and source
- Blockers and dependencies
- Decisions needed: decision, decision-maker and latest useful date
- Newly identified risks
Do not copy an item from the previous report unless the current material shows
that it remains active. Mark missing data instead of carrying a number forward.
Add a source note beside every metric and date.
[Attach source pack]
Verify: Reconcile totals, milestone dates and status labels with the systems of record. Check that completed work is not double-counted and that old blockers were not repeated automatically.
Data boundary: Limit the source pack to the project. Remove credentials, individual performance notes and customer or employee identifiers that the report does not require.
Research and decision support
6. Research a business question with sources
Use Deep Research when an answer depends on several current sources, not when a quick search will do. It can propose a research plan, work across selected websites, uploaded files and approved apps, and return a report with citations.
Provide: A precise question, date range, geography, definitions, decision context, preferred source types and exclusions.
Use: Deep Research. Restrict or prioritize websites when primary or industry-specific sources matter.
Reusable prompt:
Research this business question:
[question]
Decision this will inform: [decision]
Geography: [market]
Time period: [dates]
Definitions: [define ambiguous terms]
Prioritize: [regulators, company filings, official statistics, peer-reviewed
research, etc.]
Exclude: [source types or irrelevant topics]
Before researching, propose a plan and identify any ambiguity that could change
the answer.
The final report must include:
1. Executive answer
2. Key findings
3. Evidence table: Claim | Source | Source date | Publication type
4. Areas where reliable sources disagree
5. Unknowns and data limitations
6. Implications for the stated decision
7. Full source list
Do not present an inference as a sourced fact. Prefer original sources over
articles that merely repeat them.
Verify: Open the sources supporting the decisive claims. Check publication dates, measurement definitions and whether a citation actually supports the sentence attached to it.
Data boundary: State the decision in a sanitized way if it involves an acquisition, unreleased product, litigation, personnel matter or material non-public information.
7. Build a competitor evidence matrix
A competitor table becomes misleading when products are compared at different dates, in different currencies or through inconsistent criteria. ChatGPT can enforce a common comparison frame and make unknowns visible.
Provide: The competitors, market, customer segment, evaluation date, criteria and the decision the comparison should support.
Use: Deep Research for current public evidence. Add internal win-loss notes only if they are approved and clearly separated from public findings.
Reusable prompt:
Build an evidence-based comparison of these competitors:
[list]
Customer segment: [segment]
Market: [geography]
Evidence cutoff date: [date]
Compare on: [price, capabilities, implementation, contracts, support, security,
customer fit, etc.]
Rules:
- Use the same definition and time basis for every company.
- Normalize currencies only when an exchange-rate source and date are shown.
- Separate company claims, independently verified facts and your inferences.
- Write "not publicly confirmed" instead of estimating missing information.
Output:
1. Comparison matrix
2. Source and date for every non-trivial cell
3. Meaningful strengths and weaknesses by customer type
4. Claims that require a sales call or product test
5. Implications for [decision]
Verify: Check decisive cells against current pricing pages, documentation, filings or product tests. Watch for comparisons between list price and negotiated price, or announced features and generally available features.
Data boundary: Do not upload customer lists, confidential proposals, non-public roadmaps or information obtained under a nondisclosure agreement unless your organization has approved the use.
8. Synthesize customer interviews and survey responses
ChatGPT can code qualitative material, count recurring themes and locate counterexamples. It cannot turn a small or biased sample into representative market evidence.
Provide: Anonymized transcripts or responses, participant IDs, the research question, relevant segments and a codebook if one exists.
Use: Chat for a small set of interviews. Use Data Analysis for a structured survey export or larger coded dataset.
Reusable prompt:
Analyze these customer interviews and survey responses for:
[research question]
Use participant IDs, not names. Create an initial theme codebook, apply it
consistently and keep positive, negative and contradictory evidence.
Output:
1. Method and sample description
2. Theme table: Theme | Participants | Frequency | Supporting IDs
3. Differences by [segment], only where the sample supports comparison
4. Representative excerpts of no more than two sentences, labeled by ID
5. Counterexamples and minority views
6. Product or messaging implications
7. Questions the evidence cannot answer
Do not infer prevalence beyond this sample. Do not invent a customer quotation.
Distinguish an observed statement from your interpretation.
Verify: Recheck the most important excerpts and theme assignments against the originals. Confirm the denominator behind every percentage and make sure repeated responses from one person were not counted as separate participants.
Data boundary: Remove names, contact details, account numbers and unnecessary demographic data. Respect research consent, purpose limitations and any promise about how quotations will be used.
9. Write a decision memo with options and trade-offs
A useful decision memo exposes assumptions and alternatives instead of producing confident prose around a predetermined answer.
Provide: The decision, options already identified, constraints, evaluation criteria, evidence, deadline and who holds final authority.
Use: Chat for an interactive analysis. Use Work when the memo draws from several files and must be delivered in a set template.
Reusable prompt:
Create a decision memo for:
[decision]
Decision-maker: [role]
Deadline: [date]
Options: [list]
Constraints: [budget, time, policy, resources]
Evaluation criteria: [criteria and weights, if agreed]
Evidence: [attach or paste]
For each option:
- State the strongest case for it
- State the strongest case against it
- Identify assumptions, dependencies and irreversible consequences
- Score only against the supplied criteria and show the reasoning
Then provide:
1. Recommendation
2. Conditions under which a different option would become preferable
3. Information with the highest value before the deadline
4. Reversible next step
5. Dissenting view
Do not manufacture certainty or change the criteria to favor the recommendation.
Verify: Challenge the weights, assumptions and missing options. Have the accountable decision-maker review the evidence rather than approving the recommendation based on the summary alone.
Data boundary: Sanitize confidential strategy, deal terms, board material and personal information. Do not use ChatGPT as the final authority for legal, financial, safety or personnel decisions.
Documents, spreadsheets and presentations
10. Compare two versions of a document
This workflow is designed to answer three separate questions: what changed, what the change means and what still needs expert review.
Provide: Clearly labeled old and new versions, the document type, the sections that matter most and the purpose of the comparison.
Use: Chat for shorter text documents. Use Work for long files, several attachments or a formatted comparison report.
Reusable prompt:
Compare [OLD VERSION filename] with [NEW VERSION filename].
Purpose of review: [purpose]
Priority sections: [sections]
Create a table with:
Section | Old wording or position | New wording or position | Change type |
Practical effect | Review priority
Then list:
1. Material additions
2. Material deletions
3. Changed dates, amounts, parties, obligations or definitions
4. Internal inconsistencies introduced by the new version
5. Items whose effect is ambiguous
Use only the two supplied documents. Cite the page, heading, clause or paragraph
for every finding. Do not provide legal advice or say a change is harmless
without evidence.
Verify: Use a native document comparison or targeted manual check for high-risk clauses, tables, footnotes and tracked changes. Confirm that formatting or extraction did not hide deleted text.
Data boundary: Contracts, policies and board papers may be privileged or highly confidential. Use only an approved workspace and remove signatures, bank details, personal addresses and irrelevant exhibits when possible.
11. Review a document against a requirements checklist
Instead of asking whether a document is “good,” give ChatGPT the standard it must satisfy. This works for a proposal, grant application, tender response, policy, report or publication checklist.
Provide: The document, the authoritative requirements, the intended submission format and any rule that must be treated as mandatory.
Use: Chat for a short review. Use Work when the task includes several references, an editable revision or a formal review matrix.
Reusable prompt:
Review [document] against [requirements].
Treat the requirements as authoritative. Do not add criteria that are not in
them.
For every requirement, produce:
Requirement ID | Status: Met, Partly met, Not met or Not applicable |
Evidence location in the document | Gap | Proposed correction
Then provide:
1. Blocking issues
2. High-value improvements that are optional
3. Contradictions between the document and the requirements
4. Claims or attachments that need independent verification
5. A revised checklist for final human sign-off
If a requirement is ambiguous, quote it and explain the possible readings
without choosing one silently.
Verify: Check every “Met” judgment against the exact requirement and cited document location. Confirm formatting, word counts and required attachments in the actual submission system.
Data boundary: Remove personal identifiers from applications and submissions when they are not part of the evaluation. Do not expose confidential bids, unpublished research or protected student or client information without approval.
12. Extract obligations, owners and deadlines
ChatGPT can help build an obligation register from a contract, policy, regulation or service agreement. It should identify text for expert review, not decide what the law requires.
Provide: The source documents, relevant jurisdiction or business unit, named parties and the type of obligation you need to track.
Use: Chat or Work. Use Work when several documents must be reconciled into a finished register.
Reusable prompt:
Extract operational obligations from the supplied documents.
Focus on: [party, department or obligation type]
Create a register with:
Source | Clause or page | Responsible party | Required action | Trigger |
Deadline or frequency | Required evidence | Consequence stated in source |
Ambiguity
Rules:
- Preserve exact defined terms.
- Distinguish a fixed date from a date calculated after an event.
- Do not infer an owner when the document does not name one.
- Do not convert optional language into a mandatory obligation.
- Flag conflicts between documents.
Finish with questions for legal, compliance or operational review. This is an
extraction task, not a legal opinion.
Verify: Have the relevant legal, compliance or contract owner confirm the source clause, calculation method, exceptions and responsible party. Test several deadlines manually.
Data boundary: Use an approved environment for privileged or regulated material. Remove personal data, signatures, account details and unrelated commercial terms where they are not required.
13. Analyze a spreadsheet for trends and anomalies
ChatGPT can inspect spreadsheets and CSV files, run calculations and create tables or charts. The quality of the result depends heavily on the structure and meaning of the data.
Provide: A clean source file, a data dictionary, the business question, the reporting period, required filters and definitions for important metrics.
Use: Data Analysis. OpenAI recommends clear column names and one record per row; avoid image-based tables when exact values matter.
Reusable prompt:
Analyze the attached dataset to answer:
[business question]
Definitions:
[metric and column definitions]
Before drawing conclusions:
1. Inspect sheet names, columns, types and row counts
2. Report missing values, duplicates and impossible values
3. State the filters and exclusions you will apply
4. Ask before making a consequential assumption
Then:
- Calculate the requested metrics
- Compare [segments/time periods]
- Identify material trends and outliers
- Create [requested table/chart]
- Show the formulas or code used
- Explain limitations and alternative interpretations
Return a reconciliation table that allows the headline totals to be checked
against the source data.
Verify: Recalculate headline figures independently, inspect the code or formulas and spot-check source rows. Confirm that dates, currencies, percentages, missing values and filtered records were handled as intended.
Data boundary: Replace names and account identifiers with pseudonymous IDs. Remove payment-card information, passwords, health data and unnecessary customer or employee fields before upload.
14. Clean and restructure a messy dataset
Data cleaning becomes dangerous when changes happen silently. ChatGPT should preserve the original, propose rules and produce an exception log.
Provide: The untouched source file, intended output schema, valid value ranges, deduplication rules and the meaning of blanks.
Use: Data Analysis for tabular cleaning. Use Work when the task should produce a finished spreadsheet plus documentation.
Reusable prompt:
Prepare a cleaned copy of this dataset. Never overwrite or alter the original.
Target schema:
[columns, types and valid formats]
Cleaning rules already approved:
[rules]
First profile the data and propose any additional transformation. Do not apply
an additional rule until it is listed explicitly.
Produce:
1. Cleaned dataset
2. Transformation log: Rule | Rows affected | Example before | Example after
3. Exception file for records that cannot be resolved safely
4. Reconciliation: original rows, removed rows, merged rows and final rows
5. Data-quality summary
Do not guess missing values or merge possible duplicates without a stated rule.
Preserve source IDs so every output row remains traceable.
Verify: Compare row counts, totals and a sample of transformed values with the original. Review every destructive operation, deduplication rule and imputation before accepting the output.
Data boundary: Minimize personal data before cleaning. A messy export is not permission to upload confidential fields that the target dataset will not use.
15. Build a forecast and scenario model
ChatGPT can structure historical data, formulas and assumptions into a transparent scenario model. It cannot know which future will occur, and a polished chart does not make weak assumptions reliable.
Provide: Historical data, planning horizon, business drivers, fixed assumptions, scenario definitions and the decisions the forecast will support.
Use: Data Analysis for calculations and sensitivity testing. Use Work when the final output must be an editable spreadsheet or planning report.
Reusable prompt:
Build an assumption-based forecast for [metric/business] covering [period].
Historical data: [attach]
Known drivers: [list]
Fixed assumptions: [list]
Scenarios:
- Downside: [definition]
- Base: [definition]
- Upside: [definition]
Requirements:
1. Separate inputs, calculations and outputs
2. Show every formula and unit
3. Explain the forecasting method and why it fits the data
4. Produce monthly or quarterly results for all scenarios
5. Run sensitivity analysis on [key variables]
6. Identify break-even points and assumptions with the greatest influence
7. Include limitations and signs that the model should be updated
Do not hide uncertainty behind decimal precision. Do not invent a driver or
historical value.
Verify: Reconcile the starting period, test formulas with simple values and challenge the scenarios with subject-matter experts. Compare the model with a basic baseline such as a flat or historical-trend forecast.
Data boundary: Protect payroll, customer-level revenue, unreleased guidance and material non-public financial information. A forecast should inform an accountable decision-maker, not execute an investment, credit or staffing decision automatically.
16. Create a presentation from source material
The strongest presentation workflow starts with evidence and a communication goal, not with a request for “ten slides about” a topic.
Provide: The audience, purpose, source material, required length, brand or master template, preferred slide types and claims that must or must not appear.
Use: Work can create editable presentations and can follow supplied reference files or reusable templates where supported.
Reusable prompt:
Create an editable presentation for [audience].
Purpose: [inform, decide, approve, sell, train]
Desired action after the presentation: [action]
Length: [slides or minutes]
Required source material: [attachments]
Template or brand rules: [attachment/instructions]
First propose the narrative in one sentence per slide. After I approve it,
create the deck.
Each slide must contain:
- One clear takeaway title
- Only evidence supported by the source pack
- Source note for every external statistic or material claim
- Speaker notes explaining the reasoning
Include: [required slide types]
Exclude: [forbidden claims/topics]
Preserve the supplied master, layout, terminology and brand rules. Flag any
visual, number or quotation that requires a human rights or accuracy check.
Verify: Review every chart against its data, every claim against its source and every slide at presentation size. Check cropping, contrast, overflow, brand consistency and whether the conclusion follows from the evidence.
Data boundary: Do not upload confidential master decks, unreleased results or licensed images outside an approved environment. Confirm that you have the right to reuse source visuals, quotations and templates.
Projects and operations
17. Build a project plan
ChatGPT can turn a defined objective into work packages, milestones and dependencies. It cannot know real capacity, internal politics or hidden constraints unless you provide them.
Provide: The objective, scope, exclusions, deadline, required milestones, available roles, constraints, dependencies and definition of done.
Use: Chat for planning together. Use Work when several source documents must become a maintained project-plan artifact.
Reusable prompt:
Build a project plan for:
[objective]
In scope: [items]
Out of scope: [items]
Deadline and fixed milestones: [dates]
Available roles and realistic capacity: [details]
Known dependencies and constraints: [details]
Definition of done: [acceptance criteria]
Output:
1. Work breakdown structure
2. Milestones and deliverables
3. Dependency map and likely critical path
4. Suggested owner role for each work package
5. Decision gates and acceptance tests
6. Assumptions and unresolved questions
7. First two weeks of concrete actions
Do not assign named people or invent capacity. If the deadline conflicts with
the supplied resources, show the conflict and offer scope, time and resource
trade-offs.
Verify: Ask the people doing the work to estimate effort and confirm dependencies. Check holidays, procurement lead times, approval queues and other constraints that rarely appear in a prompt.
Data boundary: Use roles instead of employee performance data. Do not include private personnel matters, credentials, security architecture or customer secrets unless necessary and approved.
18. Create a risk register and pre-mortem
A pre-mortem asks how the project could fail before failure happens. ChatGPT is useful for broadening the list, but risk owners must judge probability, impact and controls.
Provide: The plan, objective, assumptions, environment, existing controls, risk tolerance and time horizon.
Use: Chat for facilitated brainstorming. Use Work to combine plans, audits and incident material into a formal register.
Reusable prompt:
Run a pre-mortem for [project/decision].
Imagine it is [future date] and the project failed materially. Generate plausible
failure paths across delivery, people, finance, suppliers, technology, security,
legal, compliance, customer adoption and reputation.
Then create a risk register:
Risk | Cause | Early warning | Likelihood | Impact | Existing control |
Mitigation | Contingency | Suggested owner role | Review date
Separate:
- Risks supported by the supplied material
- Plausible but unverified risks
- Unknowns that require an expert
Prioritize by expected consequence and time to intervene. Do not invent a
control or claim that a risk has been accepted.
Verify: Calibrate likelihood and impact with domain experts and historical data. Confirm that each mitigation has an owner, trigger, budget and review date.
Data boundary: Avoid exposing unpatched vulnerabilities, incident details, access paths or sensitive resilience gaps beyond the approved group and workspace.
19. Turn a process into a standard operating procedure
An SOP should make work repeatable, including exceptions and evidence. A fluent summary that omits decision points is not enough.
Provide: Current notes, process recordings or transcripts, forms, policies, roles, systems, quality controls and known exceptions.
Use: Work is well suited because it can combine several files and create an editable procedure. Use Chat when the process is short.
Reusable prompt:
Turn the supplied process material into an SOP for [role].
Purpose: [purpose]
Start trigger: [event]
End state: [completed condition]
Required structure:
1. Scope and exclusions
2. Roles and responsibilities
3. Prerequisites
4. Numbered procedure
5. Decision points with if/then paths
6. Exceptions and escalation route
7. Quality checks and required evidence
8. Systems and records used
9. Failure recovery
10. Owner, version and review date
Use the policy as authoritative if notes conflict with it. Flag every conflict
and every step that depends on undocumented knowledge. Never include a password,
secret key or reusable authentication code.
Verify: Have a person who did not help write the SOP perform a dry run. Confirm permissions, screen labels, exception routes, control evidence and the document owner.
Data boundary: Replace live credentials and customer records with placeholders or synthetic examples. Limit security procedures and internal system details to the approved audience.
Marketing, support and people
20. Draft a marketing campaign brief
ChatGPT can organize evidence into a coherent campaign and generate testable message options. It should not invent product claims, customer proof or legal approval.
Provide: The product facts, audience research, offer, funnel stage, channel, budget range, brand voice, proof points, forbidden claims and success metric.
Use: Chat for development and critique. Use Work when research, a template and several assets must become one campaign package.
Reusable prompt:
Create a campaign brief for [product/offer].
Audience: [segment and evidence]
Business objective: [objective]
Desired customer action: [action]
Channel and funnel stage: [details]
Approved proof points: [facts and sources]
Brand voice: [rules/examples]
Claims we may not make: [list]
Budget and timing: [constraints]
Produce:
1. Audience problem and insight
2. Single-minded proposition
3. Message hierarchy
4. Three creative territories
5. Required assets by channel
6. Test plan with one variable per test
7. Primary and guardrail metrics
8. Approval and compliance checklist
Use only approved facts. Label hypotheses as hypotheses. Do not fabricate a
testimonial, performance result, scarcity claim or customer statistic.
Verify: Check every product claim, price, customer quote, link and legal disclosure. Ensure the test design can distinguish what caused a result.
Data boundary: Do not paste customer lists, ad-platform credentials, unpublished launch details or personally identifiable targeting data into an unapproved account.
21. Repurpose one approved source into several channels
Repurposing should adapt the same verified idea to different contexts. It should not multiply unsupported claims across five platforms.
Provide: The approved source, brand guide, intended audiences, channel requirements, calls to action and links.
Use: Chat for a small asset set. Use Work when the task includes several files, media formats or a finished campaign package.
Reusable prompt:
Repurpose the approved source below into:
[list of channels and asset types]
Audience by channel: [details]
Voice rules: [rules]
Required call to action: [CTA]
Approved links: [links]
Rules:
- Treat the source as the only factual authority.
- Preserve qualifications and uncertainty.
- Adapt the framing and length to each channel; do not merely truncate.
- Do not add a statistic, quotation, product feature or promise.
- Show which source section supports each asset.
For every asset, provide:
1. Final copy
2. Character or word count
3. Source mapping
4. Suggested visual concept, if relevant
5. Claim or link requiring final review
Verify: Compare every asset with the approved source, check channel limits and test all links. Make sure a shortened version did not remove a necessary caveat.
Data boundary: Do not upload embargoed announcements, licensed source material or private customer stories unless reuse in the selected channels has been approved.
22. Draft customer-support replies grounded in policy
ChatGPT is helpful when the reply must be empathetic and consistent with an approved policy. It should not decide entitlements or issue refunds from an incomplete account history.
Provide: The customer’s message, relevant account facts, the current policy or knowledge-base article, permitted remedies and escalation rules.
Use: Chat for a draft based on pasted material. Use Work or an approved support app when retrieving case history or saving a draft. Keep external sending and irreversible account actions subject to review.
Reusable prompt:
Draft a customer-support reply using only the customer message, verified account
facts and approved policy below.
Goal: [resolve, request information, explain, escalate]
Permitted remedies: [list]
Required disclosure: [text]
Tone: calm, specific and respectful
Output:
1. Draft reply
2. Policy rule supporting the response
3. Missing account fact that could change the answer
4. Recommended internal next step
Do not blame the customer. Do not promise a refund, deadline, exception or
technical outcome unless it is explicitly authorized. If the policy and account
facts do not determine the answer, draft an escalation instead.
Do not send or change the account.
Verify: Confirm identity and entitlement through the authorized system, check policy version and inspect any promised date or remedy. A human should approve exceptions and high-impact account actions.
Data boundary: Never request or include a password, full payment-card number, authentication code or unnecessary identity document. Minimize order, account and health-related information.
23. Find root causes and knowledge gaps in support tickets
Anonymized ticket analysis can reveal recurring failure modes, confusing product behavior and missing documentation. Frequency alone does not prove root cause or business impact.
Provide: An anonymized ticket export, date range, product taxonomy, severity definitions, channel and known changes during the period.
Use: Data Analysis for classification, counting, trend analysis and charts.
Reusable prompt:
Analyze the attached support-ticket export covering [dates].
First report the schema, missing data, duplicates and how one ticket is defined.
Then create and apply a reproducible issue taxonomy.
Output:
1. Ticket volume by issue, severity, product area and week
2. Emerging or declining issues
3. Repeat contacts and unresolved cases, if the data supports them
4. Suspected root causes, clearly separated from observed evidence
5. Product, process or documentation gaps
6. Ten representative anonymized ticket IDs for manual review
7. Recommended next investigation and expected evidence
Show classification rules and confidence. Do not infer customer intent or a
technical root cause from message wording alone.
Verify: Manually review samples from the largest and highest-severity categories. Check duplicate handling, changes in ticket volume, channel mix and releases that could distort the trend.
Data boundary: Remove names, emails, phone numbers, account identifiers, message signatures and sensitive free-text details. Use case IDs that authorized reviewers can map back inside the support system.
24. Standardize a hiring process without delegating the decision
ChatGPT can translate job requirements into a scorecard, structured interview questions and consistent evidence notes. It should not infer suitability from protected characteristics or make the final employment decision.
Provide: Essential job outcomes, job-related competencies, approved job-description language, interview stages, legal or policy constraints and the existing scorecard.
Use: Chat for the framework. Use Work when several policy documents and templates must become a hiring kit.
Reusable prompt:
Create a structured hiring kit for [role].
Role outcomes after 6 and 12 months: [outcomes]
Essential job-related competencies: [list]
Approved requirements: [list]
Interview stages: [stages]
Policy or legal constraints: [details]
Produce:
1. Clear job description using only approved requirements
2. Scorecard with observable, job-related evidence for each criterion
3. Structured interview questions asked consistently to every candidate
4. Follow-up probes and strong/weak evidence indicators
5. Interview note template: Evidence | Criterion | Uncertainty
6. Work-sample exercise and scoring guide
7. Human review and accommodation checkpoints
Do not infer or use age, race, ethnicity, nationality, religion, sex, gender,
sexual orientation, disability, health, family status or other protected or
irrelevant personal characteristics. Do not recommend a hire.
Verify: Have HR or legal reviewers confirm the requirements, questions, scoring and accommodations. Train interviewers to record evidence rather than impressions and retain human accountability for every decision.
Data boundary: Do not upload resumes or notes containing unnecessary photos, addresses, birth dates, government IDs, health data, family information or other protected details. Follow applicable employment and retention rules.
Recurring work
25. Schedule a briefing or change monitor
Scheduled Tasks can run once, repeat or periodically check for a meaningful change. The best task defines both what deserves a notification and what should produce silence.
Provide: The subject, approved sources, timing, timezone, materiality threshold, required output and stop condition.
Use: Scheduled Tasks. An approved connected app can supply information when the task has permission to read it. A task created inside a Project currently cannot use that Project’s files, so include the necessary baseline in the task itself or place it in an accessible approved source.
Reusable prompt:
Check [subject] every [schedule] in [timezone].
Use only: [approved websites or connected sources]
Baseline: [current state or comparison date]
A meaningful change is: [specific threshold or condition]
When a meaningful change occurs, send:
1. What changed
2. When it changed
3. Source link or record
4. Difference from the baseline
5. Why it may matter
6. What should be checked next
If nothing meets the threshold, do not notify me.
Do not take any external action or change a connected record.
Stop after [date/condition], if applicable.
Verify: Review the first few runs for noise, missing sources and false positives. Make the threshold more specific when notifications do not lead to an action. Periodically confirm that the source and baseline are still current.
Data boundary: Give connected apps the least access the task requires. Keep monitoring read-only unless a separately reviewed workflow genuinely requires an action, and do not place confidential baselines in a personal task.
How to turn a useful prompt into a reliable workflow
Do not automate the first result. Stabilize the work first.
1. Start with low-risk, frequent work
Choose a task that happens often, uses accessible source material and has an output a person can check quickly. Weekly reports and document checklists are usually easier starting points than legal interpretation or employee assessment.
2. Collect five real examples
Run the workflow on several representative cases, including at least one messy or difficult example. A prompt that succeeds on a perfect file may fail when columns move, evidence conflicts or the writer omits a key fact.
3. Define acceptance criteria
Decide what makes the result usable before running it. Criteria may include correct totals, every claim linked to a source, no unsupported commitments, a maximum word count or zero unresolved required fields.
4. Preserve the source and the output
Keep the original inputs, prompt version, generated output and human corrections where policy allows. This makes errors diagnosable and shows whether the workflow is improving.
5. Create an escalation path
Tell users when to stop. Examples include a missing source, a conflict between policies, a low-confidence classification, a request for legal interpretation or an external action that requires authority.
6. Standardize only after review
Once the workflow performs consistently, place the files and instructions in a Project, reuse an approved template or assign the recurring portion to a Scheduled Task. Do not remove review merely because the output looks professional.
How to measure whether ChatGPT is helping
Measure the complete process, including preparation and correction.
| Metric | What to measure |
| Cycle time | Minutes from receiving the inputs to an approved output |
| First-pass acceptance | Percentage requiring only minor edits |
| Correction load | Number and severity of factual, calculation and policy corrections |
| Coverage | Required items found versus items missed |
| Source accuracy | Claims whose cited source actually supports them |
| Rework | Work repeated because the first result was unusable |
| Risk events | Sensitive-data, permission, compliance or external-action failures |
| User value | Whether the output changed or accelerated a real decision |
Compare the AI-assisted workflow with the existing process on similar cases. A draft that appears in 30 seconds but takes an hour to verify is not necessarily an improvement. A workflow that saves ten minutes while increasing error risk may be worse.
Work that should keep a human decision-maker
ChatGPT can prepare evidence, alternatives and drafts. It should not be the final authority for:
- sending an important external communication;
- publishing an unverified factual claim;
- signing, accepting or interpreting a contract;
- hiring, firing, promotion or employee discipline;
- medical, legal, credit, investment or safety decisions;
- changing access controls or sharing permissions;
- deleting records or executing financial transactions;
- approving work where the reviewer cannot inspect the underlying evidence.
The dividing line is not whether ChatGPT sounds confident. It is whether the action affects rights, money, safety, access, reputation or another person in a way that is difficult to reverse.
Frequently asked questions
How can I use ChatGPT for work?
Start with a specific task such as drafting an email, preparing a meeting brief, comparing documents, analyzing a spreadsheet or producing a status report. Provide the authoritative source material, define the required output and ask ChatGPT to show missing information and verification points. Review the result before using it.
What is the best ChatGPT workflow for beginners?
A low-risk, recurring task with a clear answer is the best place to start. Try converting notes into a fixed report format or checking a document against a checklist. Avoid starting with a confidential, irreversible or high-stakes decision.
Do I need ChatGPT Work to use ChatGPT at work?
No. Regular Chat can handle many short professional tasks. ChatGPT Work is the named experience for longer, multi-step assignments and finished deliverables. Deep Research, Data Analysis, Projects, apps and Scheduled Tasks solve different parts of a workplace process, depending on availability.
Is ChatGPT Work the same as ChatGPT Business?
No. ChatGPT Work is a way of completing longer tasks inside ChatGPT. ChatGPT Business is a paid organizational workspace with team administration and business data protections. Work may be available through several paid plans. See our ChatGPT pricing comparison for the current plan differences.
Can I use ChatGPT for work for free?
Many basic drafting, summarizing, search and file tasks can be tested on the Free plan, but limits and access to advanced models or tools differ. Start with the real task, observe the constraint and then compare subscription options rather than assuming a paid tier is necessary.
Is it safe to upload work documents to ChatGPT?
That depends on the document, account, workspace, settings, contracts, law and employer policy. Use an approved environment, minimize data, redact unnecessary identifiers and never upload secrets or credentials. OpenAI says business workspace data is not used for model training by default, while personal users can manage model-improvement settings through Data Controls. Retention and permission questions remain separate from training.
Can ChatGPT analyze Excel spreadsheets?
ChatGPT can analyze common spreadsheet and data formats, calculate metrics and create tables or charts. For reliable results, use descriptive headers, one record per row and clear metric definitions. Review the generated formulas or code and independently reconcile important totals.
Can ChatGPT create reports and presentations?
Yes. ChatGPT Work can create or edit documents, spreadsheets, presentations and reports where supported. Supply the purpose, audience, source material, output format and any template or brand rules. Always inspect the finished file rather than judging only the chat response.
Can ChatGPT read my work email, calendar or company files?
ChatGPT can access supported external services only through an app that is connected, authorized and allowed by the workspace. Available read and write actions depend on the app, plan, role and administrator settings. Review the requested permission and each consequential action before approval.
Can ChatGPT automate recurring work?
Scheduled Tasks can run one-off or recurring work and monitor for changes. Define a schedule, source, threshold, output and stop condition. Tasks cannot run more often than hourly, and current tool, plan and Project-file limitations still apply.
How do I stop ChatGPT from inventing facts at work?
You cannot guarantee that it will never make an error. Reduce the risk by supplying authoritative sources, forbidding unsupported claims, requiring source locations, making unknowns explicit and verifying the decisive facts yourself. For calculations, inspect the method and reconcile totals with the original data.
Should a company use employees’ personal ChatGPT accounts?
That is a governance decision. An organization should consider approved data handling, administration, offboarding, retention, app access and workspace controls rather than reimbursing personal subscriptions without a policy. Business and Enterprise plans address organizational needs that are separate from an individual model allowance.
The bottom line
Using ChatGPT well at work is less about discovering one perfect prompt and more about designing a controlled handoff between people, evidence and software.
Give ChatGPT the smallest reliable source pack that can support the task. Define the deliverable and the rules. Ask it to expose missing information rather than smoothing over the gaps. Then inspect the evidence, correct the result and keep a person accountable for the final action.
Start with one frequent, low-risk workflow from this guide. Run it on five real examples. Measure total time and correction load. Only then turn it into a shared Project, template or Scheduled Task.
For the broader product explanation, continue with our complete ChatGPT guide. To choose a plan based on features, limits and team controls, use our ChatGPT pricing guide. You can also explore the latest ChatGPT news and guides and learn more about OpenAI.
Popular news
Best ChatGPT Alternatives in 2026: Claude, Gemini, Copilot and More
ChatGPT: Complete Guide to Features, Models, Pricing and How It Works
EU slaps Google with €890 million fine for breaching the Digital Markets Act
Google Rebrands NotebookLM as Gemini Notebook
Is ChatGPT Safe? Privacy, Security, Training Data and Business Use
ChatGPT Pricing Explained: Free vs Go, Plus, Pro, Business and Enterprise
Latest comments
Loading